Enchantment with Computer Reason

Today, when computer systems are so ubiquitous and therefore mundane, so immensely powerful but yet taken for granted, how do programmers motivate themselves? How do they get out of bed in the mornings and say, “Yep! I can’t wait to get on the keyboard!”

There was a time when this motivation seemed easy to explain. Think of Sherry Turkle’s book, The Second Self. There she described how exciting it was to get a machine to ‘act’ in accord with one’s own instructions. She said that getting one’s thoughts onto a screen and for those thoughts to get the machine to function like one’s second self was magical, something that really motivated you. But her book was written when the machines on the desk being one’s second self were called microcomputers, not personal computers. That gives an idea of how long ago that excitement was. So today – what enchants coders?

I do think it is enchantment that gets the coder out of bed but I think this is a quite different kind from that which Turkle described. Indeed it is almost reverse, one might say. In my view, many coders today find their enchantment in Machine Learning. They are enchanted because machine learning makes computers act in ways that they, the coder, cannot understand. It is not their reasoning writ large on the performance of the machine that excites them or provokes a sense of wonder; it is, on the contrary, how the machine works despite them that is.

The aspect of computer programming I am thinking of is a part of machine learning that is sometimes called Deep Learning. This is part of a broader family of methods based on the notion that programmes themselves can, as it were, ‘learn’ how to correctly represent data and thus act on that data. In the approach I am thinking of, no human is required to label data as part of some training set. Rather, the machine or rather the application somehow ‘uncovers’ categories and features in the data (about the world, say) and then acts accordingly.

What comes to mind, particularly, are computer vision systems, where certain programmes are able to identify (to ‘see’, as it were) objects not merely as a function of ‘unsupervised learning’, a technique whereby the programmes come to recognise objects without the aid of a human expert, for such techniques presuppose that what the system finds accords with what the human programmer can see too – the machine in this sense is only copying what the human can do, though doing this autonomously. In contrast, these new systems are identifying objects – patterns, shapes, phenomena in the visual field – that no human could see. They are, if you like, doing something beyond what the human can do.

As it happens, and in many instances, various advanced computer vision processing applications have been doing this for some time – though without the fanfare that has erupted recently.

Good examples of what such programmes can do can be found in the work of, for example, Graham Budgett, an artist at the University of California, Santa Barbara. Here, the images he produces, his art if you like, are to be seen through a browser. These images keep iterating and changing themselves as you look. They do so as a function of the algorithms that make the images you see a transitory output. That is to say, these algorithms constantly reinterpret the objects, the shapes, the forms, the colours, that Budgett provides for them in the first place. The algorithms present these as the first thing one sees. But then they start interpreting and reinterpreting these shapes, colours, forms. In each cycle of interpretation, the code starts with a same initial set of objects (whatever they might be), and these are processed and interpreted results in infinitely new forms every time the code (or the application) is run. The code is probalistic, not deterministic, and so comes up with different interpretations each time it parses.

In a sense on might say that the art here – the painting if you prefer but no paintbrushes are involved, only a keyboard and mouse – is being done by code. What the artist does, in this case Budgett, is select the machine learning algorithms as if they were paints for the palette. The ‘art’ comes to be in how the code interacts with its own output; thus Budgett has created art that performs without his controlling hand.

Though his examples are only of pictorial art, in important respects the pictures are showing something quite radical. The applications producing these pictures are not articulating human knowledge, knowledge about shapes and objects in the world. Rather, they are creating, through the application’s interpretation, new knowledge, new forms and shapes. These are the dynamic output from algorithms. In these respects, the Turing Test has been passed in radically impressive way since computing is not so much mimicking human intelligence, as it is doing something people cannot do – making thing with a new kind of intelligence.

This is significant. If this is the enchantment that coders are finding today, then, this is fundamentally different to the kind described by Turkle in Second Self. If, then, the delight she described was in getting a machine to act according to a coder’s own reasons, now the delight that coders feel is in getting machines to act in terms of reasons that the machine produces. The enchantment is no longer in the self, in how one gets a machine to act as a mirror of one’s own thoughts; it is in how some application can reason autonomously. It is as if the coders want the applications they code to do something more than the coders can imagine themselves.

Now for many coders this seems to be an enchanting moment. Here at last is a glimpse of what they have been seeking since the term ‘AI’ was first made common currency after the Dartmouth Conference where the term was first coined in 1956.

The trick, though, is that the applications that are currently being sought are ones that seem to have reasons that people don’t have, that people couldn’t have, that are more than human in their intelligence. And here it is not simply that computers can process at vast speed, that they are simply better calculators; on the contrary, the coders think that the applications they are producing reason in ways that is beyond human reason.

This is somehow beyond what Turing imagined. Given the deity like status this mathematician has in the pantheon of computer science, this is presumably enormously exciting to the coder. No wonder they are so keen to get out of bed. It’s not what they do that excites them, its what their applications will do.

Advertisement